[Mlir-commits] [mlir] Extend NVVM mma ops to support fp64 (PR #165380)

Giacomo Castiglioni llvmlistbot at llvm.org
Fri Oct 31 02:55:30 PDT 2025


================
@@ -954,6 +960,14 @@ LogicalResult NVVM::WMMALoadOp::verify() {
     return emitOpError() << "invalid attribute combination";
   std::pair<Type, unsigned> typeInfo = inferMMATypeFromMNK(
       getEltype(), getFrag(), getM(), getN(), getK(), getContext());
+  // Special case for f64 fragments
+  Type f64Ty = Float64Type::get(getContext());
+  if (typeInfo.first == f64Ty && typeInfo.second == 1) {
+    if (getType() != f64Ty)
+      return emitOpError("expected destination type to be f64");
----------------
castigli wrote:

Done.

https://github.com/llvm/llvm-project/pull/165380


More information about the Mlir-commits mailing list